01-05-2015 11:28 AM
You may need to define your meaning for "accuracy" for this purpose.
And the answer could well depend on the data you are testing. If the data is binomial in nature having more than 2 classes is likely to be difficult to create and/or interpret.
03-28-2015 06:01 AM
I think a few of us are confused with the terms you've used in your original question. Reading further, it appears to me that you might be asking about the statistical power of a likelihood ratio of wald test, and how that is related to the number of discrete levels in the categorical variable that is the focus of the test.
If that is your underlying concern, increasing or decreasing the number of levels in a categorical predictor variable may or may not increase the overall fit of the model, and may or may not improve any overall model fit statistics (not that many, if any, of those have a known distribution that leads to useful tests), and can often decrease some of them (the ones that use degrees of freedom of the model to penalise the overall fit statistic). Overall fit statistics would be improved if the explanatory power of a predictor recoded with more levels was better than that of the predictor coded with fewer levels, but that does not always happen.
Does that help? Did I guess correctly?
01-27-2015 01:03 AM
It depends on the number of observations being tested....should always have at least 30 samples (at least 1000 samples for research) especially if you want accurate results from the chi-square test; this way the expected numbers for each classes are more than 5. If the samples are less than above then you should probably try pooling those classes with small expected numbers or use an Exact Test method.